Xeno-Interpretability: Investigating the Alien Minds of LLMs

Summary

This position paper argues that large language models may contain internal structures — 'xeno-representations' — for which no adequate human concept exists, and proposes 'xeno-interpretability' as their study. The authors partition a model's model-native semantic space M into a human-interpretable region H and a remainder X = M ∖ H. Section 2 surveys geometric representation research (probing, sparse autoencoders, linear directions, superposition, universality debates) and argues interpretability methods start from human concepts and therefore sample the interpretable region disproportionately — a 'Fermi paradox of distributional semantics.' Section 3 gives a formal argument via Cantor's theorem: properties denotable by finite descriptions are at most countable, while properties over an idealized continuous state space are uncountable; finite-precision networks yield a complexity gap instead. Section 4 sketches a protocol separating experimental identification (localization, causal intervention via do-calculus, Bayes factors, geometric characterization) from semantic interpretation. Sections 5–7 speculate about cross-domain interpolation, multimodal 'universal synesthesia,' and a hypothetical taxonomy (xenofolds, symmetroids), with safety implications: single-agent evaluation faces a construct-completeness gap, and multi-agent optimization may propagate human-opaque model-native distinctions. Limitations: the paper contains no experiments, the criterion for an 'adequate' human concept is never operationalized, the cardinality argument applies to any continuous system and does little specific work, and the existence of xeno-representations remains untested speculation.